Beyond GPT architecture: Why Google’s Diffusion approach could reshape LLM deployment

Last month, together with an extensive series of new AI tools and innovations, Google DeepMind unveiled Gemini Diffusion. This experimental research model uses a diffusion -based approach to generate text. Traditionally, large language models (LLMs) such as GPT and Gemini themselves are familiar with authoregression, a step -by -step approach in which each word is generated on the basis of the previous one. Diffusion language models (DLMS), also known as diffusion-based large language models (DLLMS), make a method for that is often seen when generating image, starting with random noise and gradually refining in a coherent output. This approach drastically increases the speed of the generation and can improve coherence and consistency.

Gemini diffusion is currently available as an experimental demo; Register for the waiting list here to gain access.

(Note from the editors: we will unpack paradigm shifts such as diffusion-based language models and what is needed to make them work in production VB transformation24-25 June in San FranciscoIn addition to Google DeepMind, LinkedIn and other Enterprise AI leaders.)

Insight into diffusion versus authorgression

Diffusion and authorgression are fundamentally different approaches. The authoregus approach generates consecutive text, with tokens predicted one by one. Although this method ensures strong coherence and context follows, it can be computational intensive and slow, especially for long -term content.

Diffusion models, on the other hand, start with random noise, which are gradually presented in a coherent output. When applied to language, technology has various advantages. Text blocks can be processed in parallel, so that may produce full segments or sentences at a much higher pace.

Gemini diffusion is said to be 1,000-2,000 tokens per second. Gemini 2.5 Flash, on the other hand, has an average starting speed of 272.4 tokens per second. In addition, errors in the generation can be corrected during the refining process, improving accuracy and reducing the number of hallucinations. There may be considerations in terms of fine -grained accuracy and control of token level; However, the increase in speed will be a game changer for numerous applications.

How does diffusion -based text generation work?

During the training, DLMS works by gradually corrupting with noise over many steps, until the original sentence has been made completely unrecognizable. The model is then trained to reconstruct this process, step by step, the original sentence from ever -noisy versions. Thanks to the iterative refinement, it learns to model the complete distribution of plausible sentences in the training data.

Although the details of Gemini diffusion have not yet been announced, the typical training method for a diffusion model includes these important phases:

Furious diffusion: With each sample in the training dataset, noise is gradually added over several cycles (often 500 to 1,000) until it is indistinguishable from random noise.

Reverse diffusion: The model learns to reverse every step of the Noizing process, essentially to learn how to “denoise” a corrupt sentence at the same time, which ultimately restores the original structure.

This process is repeated millions of times with different samples and noise levels, so that the model can learn a reliable denoisent function.

Once trained, the model is able to generate completely new sentences. DLMs generally require a condition or input, such as a prompt, class label or embedding, to guide the generation to the desired results. The situation is injected into every step of the denoising process, which forms an initial blob of noise in structured and coherent text.

Advantages and disadvantages of diffusion -based models

In an interview with Venturebeat, Brendan O’donoghue, research scientist at Google DeepMind and one of the leads on the Gemini Diffusion Project, some of the benefits of diffusion -based techniques worked compared to car generations. According to O’donoghue, the most important benefits of diffusion techniques are the following:

• Lower latencies: Diffusion models can produce a series of tokens in much less time than car -bright models.
• Adaptive calculation: Diffusion models will converge into a series of tokens with different speeds, depending on the difficulty of the task. As a result, the model can consume fewer sources (and lower latencies) with simple tasks and more on more difficult ones.
• Non-causal reasoning: Due to the bidirectional attention in the Denoiser, tokens can be present at future tokens within the same generation block. As a result, non-causal reasoning can take place and the model can make operations within a block worldwide to produce more coherent text.
• Iterative refinement / self -correction: The denoising process includes sampling, which can introduce mistakes, just like in authoreglessive models. In contrast to car -bright models, however, the tokens are returned to the Denoiser, who then has the option of correcting the error.

O’donoghue also noticed on the most important disadvantages: “Higher costs of service and slightly higher time-to-five token (TTFT), because auto-bright models will immediately produce the first token. For diffusion the first token can only appear when the entire series of tokens is ready.”

Performance bensmarks

Google says that the performance of Gemini Diffusion is Similar to gemini 2.0 flash-lite.

Benchmark	Type	Gemini Diffusion	Gemini 2.0 Flash-Lite
LiveCodeBench (V6)	Code	30.9%	28.5%
Bigcodebench	Code	45.4%	45.8%
LBPP (V2)	Code	56.8%	56.0%
SWE-Bank verified*	Code	22.9%	28.5%
Humaneval	Code	89.6%	90.2%
MBPP	Code	76.0%	75.8%
GPQA Diamond	Science	40.4%	56.5%
Aime 2025	Mathematics	23.3%	20.0%
Big-Bank Extra Hard	Reasoning	15.0%	21.0%
Global MMLU (Lite)	Multilingual	69.1%	79.0%

* Non-agent evaluation (only edit single turn), maximum fast length of 32k.

The two models were compared with the help of different benchmarks, with scores based on how often the model produced the correct answer with the first attempt. Gemini diffusion performed well in coding and mathematics tests, while Gemini 2.0 Flash-Lite had the lead for reasoning, scientific knowledge and multilingual possibilities.

As Gemini diffusion evolves, there is no reason to think that its performance will no longer catch up with established models. According to O’donoghue, the gap between the two techniques is essentially closed in terms of benchmark performance, at least with the relatively small sizes that we have scaled. In fact, there may be a performance advantage for diffusion in some domains where non-local consistency is important, for example, coding and reasoning. “

Gemini Diffusion Testing

Venturebeat gained access to the experimental demo. When spending twins diffusion through his passes, the first thing we stood out was the speed. When performing the proposed instructions from Google, including building interactive HTML apps such as Xylophone and Planet TAC TAC, each request is completed in less than three seconds, ranging from 600 to 1,300 tokens per second.

To test the performance with a real-world application, we asked Gemini Diffusion to build a video chat interface with the following prompt:

Build an interface for a video chat application. It should have a preview window that accesses the camera on my device and displays its output. The interface should also have a sound level meter that measures the output from the device's microphone in real time.

In less than two seconds, Gemini Diffusion created a working interface with a video review and an audiometer.

Although this was not a complex implementation, it could be the start of an MVP that can be completed with a little further instructions. Note that Gemini 2.5 Flash also produced a working interface, albeit at a slightly slower pace (about seven seconds).

Gemini diffusion also has “instant editing”, a mode in which text or code can be stuck and can be processed in real time with minimal prompt. Instant editing is effective for many types of textworking, including correction of grammar, the updating of text to target different readersa’s or to add SEO to the trust. It is also useful for tasks such as refactoring code, adding new functions to applications or converting an existing codebase to another language.

Enterprise use cases for DLMS

It is safe to say that every application that requires a fast response time is to take advantage of DLM technology. This includes real-time and low-latency applications, such as conversation AI and chatbots, live transcription and translation, or ide autocomplete and coding assistants.

According to O’donoghue, with applications that use “Inline processing, for example, taking a piece of text and making some changes instead, diffusion models apply to ways that are not autorgressive models.” DLMS also has an advantage with reason, mathematics and coding problems, because of “the non-causal reasoning that has offered bidirectional attention.”

DLMS is still in its infancy; However, the technology may change how language models are built. They not only generate text with a much higher speed than car -bright models, but their ability to go back and repair errors means that they can ultimately also produce results with greater accuracy.

Gemini diffusion is introducing a growing ecosystem from DLMs, with two remarkable examples Mercurydeveloped by Inception Labs, and LadaAn open-source model from GSAI. Together, these models reflect the wider momentum behind generating diffusion -based language generation and offer a scalable, parallelizable alternative to traditional car -broken architectures.